2 * Copyright (C) 2017-2019 Alyssa Rosenzweig
3 * Copyright (C) 2017-2019 Connor Abbott
4 * Copyright (C) 2019 Collabora, Ltd.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
26 #include <midgard_pack.h>
34 #include "util/macros.h"
35 #include "util/u_math.h"
37 #include "midgard/disassemble.h"
38 #include "bifrost/disassemble.h"
40 #include "pan_encoder.h"
42 static void pandecode_swizzle(unsigned swizzle
, enum mali_format format
);
44 #define MEMORY_PROP(obj, p) {\
46 char *a = pointer_as_memory_reference(obj->p); \
47 pandecode_prop("%s = %s", #p, a); \
52 #define MEMORY_PROP_DIR(obj, p) {\
54 char *a = pointer_as_memory_reference(obj.p); \
55 pandecode_prop("%s = %s", #p, a); \
60 #define DUMP_CL(title, T, cl, indent) {\
61 fprintf(pandecode_dump_stream, "%s\n", title); \
62 struct MALI_ ## T temp; \
63 MALI_ ## T ## _unpack((const uint8_t *) cl, &temp); \
64 MALI_ ## T ## _print(pandecode_dump_stream, &temp, indent * 2); \
67 #define MAP_ADDR(T, addr, cl) \
68 const uint8_t *cl = 0; \
70 struct pandecode_mapped_memory *mapped_mem = pandecode_find_mapped_gpu_mem_containing(addr); \
71 cl = pandecode_fetch_gpu_mem(mapped_mem, addr, MALI_ ## T ## _LENGTH); \
74 #define DUMP_ADDR(title, T, addr, indent) {\
75 MAP_ADDR(T, addr, cl) \
76 DUMP_CL(title, T, cl, indent); \
79 FILE *pandecode_dump_stream
;
81 /* Semantic logging type.
83 * Raw: for raw messages to be printed as is.
84 * Message: for helpful information to be commented out in replays.
85 * Property: for properties of a struct
87 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
90 enum pandecode_log_type
{
96 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
97 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
98 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
100 unsigned pandecode_indent
= 0;
103 pandecode_make_indent(void)
105 for (unsigned i
= 0; i
< pandecode_indent
; ++i
)
106 fprintf(pandecode_dump_stream
, " ");
110 pandecode_log_typed(enum pandecode_log_type type
, const char *format
, ...)
114 pandecode_make_indent();
116 if (type
== PANDECODE_MESSAGE
)
117 fprintf(pandecode_dump_stream
, "// ");
118 else if (type
== PANDECODE_PROPERTY
)
119 fprintf(pandecode_dump_stream
, ".");
121 va_start(ap
, format
);
122 vfprintf(pandecode_dump_stream
, format
, ap
);
125 if (type
== PANDECODE_PROPERTY
)
126 fprintf(pandecode_dump_stream
, ",\n");
130 pandecode_log_cont(const char *format
, ...)
134 va_start(ap
, format
);
135 vfprintf(pandecode_dump_stream
, format
, ap
);
139 /* To check for memory safety issues, validates that the given pointer in GPU
140 * memory is valid, containing at least sz bytes. The goal is to eliminate
141 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
142 * overruns) by statically validating pointers.
146 pandecode_validate_buffer(mali_ptr addr
, size_t sz
)
149 pandecode_msg("XXX: null pointer deref");
155 struct pandecode_mapped_memory
*bo
=
156 pandecode_find_mapped_gpu_mem_containing(addr
);
159 pandecode_msg("XXX: invalid memory dereference\n");
165 unsigned offset
= addr
- bo
->gpu_va
;
166 unsigned total
= offset
+ sz
;
168 if (total
> bo
->length
) {
169 pandecode_msg("XXX: buffer overrun. "
170 "Chunk of size %zu at offset %d in buffer of size %zu. "
171 "Overrun by %zu bytes. \n",
172 sz
, offset
, bo
->length
, total
- bo
->length
);
177 struct pandecode_flag_info
{
183 pandecode_log_decoded_flags(const struct pandecode_flag_info
*flag_info
,
186 bool decodable_flags_found
= false;
188 for (int i
= 0; flag_info
[i
].name
; i
++) {
189 if ((flags
& flag_info
[i
].flag
) != flag_info
[i
].flag
)
192 if (!decodable_flags_found
) {
193 decodable_flags_found
= true;
195 pandecode_log_cont(" | ");
198 pandecode_log_cont("%s", flag_info
[i
].name
);
200 flags
&= ~flag_info
[i
].flag
;
203 if (decodable_flags_found
) {
205 pandecode_log_cont(" | 0x%" PRIx64
, flags
);
207 pandecode_log_cont("0x%" PRIx64
, flags
);
211 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
212 static const struct pandecode_flag_info gl_enable_flag_info
[] = {
213 FLAG_INFO(OCCLUSION_QUERY
),
214 FLAG_INFO(OCCLUSION_PRECISE
),
215 FLAG_INFO(FRONT_CCW_TOP
),
216 FLAG_INFO(CULL_FACE_FRONT
),
217 FLAG_INFO(CULL_FACE_BACK
),
222 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
223 static const struct pandecode_flag_info clear_flag_info
[] = {
226 FLAG_INFO(SLOW_STENCIL
),
231 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
232 static const struct pandecode_flag_info mask_flag_info
[] = {
241 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
242 static const struct pandecode_flag_info u3_flag_info
[] = {
244 FLAG_INFO(PER_SAMPLE
),
245 FLAG_INFO(CAN_DISCARD
),
246 FLAG_INFO(HAS_BLEND_SHADER
),
247 FLAG_INFO(DEPTH_WRITEMASK
),
248 FLAG_INFO(DEPTH_CLIP_NEAR
),
249 FLAG_INFO(DEPTH_CLIP_FAR
),
253 static const struct pandecode_flag_info u4_flag_info
[] = {
255 FLAG_INFO(NO_DITHER
),
256 FLAG_INFO(DEPTH_RANGE_A
),
257 FLAG_INFO(DEPTH_RANGE_B
),
258 FLAG_INFO(STENCIL_TEST
),
259 FLAG_INFO(ALPHA_TO_COVERAGE
),
264 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
265 static const struct pandecode_flag_info mfbd_fmt_flag_info
[] = {
271 #define FLAG_INFO(flag) { MALI_AFBC_##flag, "MALI_AFBC_" #flag }
272 static const struct pandecode_flag_info afbc_fmt_flag_info
[] = {
278 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
279 static const struct pandecode_flag_info mfbd_extra_flag_hi_info
[] = {
285 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
286 static const struct pandecode_flag_info mfbd_extra_flag_lo_info
[] = {
292 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
293 static const struct pandecode_flag_info shader_midgard1_flag_lo_info
[] = {
296 FLAG_INFO(READS_TILEBUFFER
),
297 FLAG_INFO(WRITES_GLOBAL
),
302 static const struct pandecode_flag_info shader_midgard1_flag_hi_info
[] = {
304 FLAG_INFO(SUPPRESS_INF_NAN
),
309 #define FLAG_INFO(flag) { MALI_BIFROST_##flag, "MALI_BIFROST_" #flag }
310 static const struct pandecode_flag_info shader_bifrost_info
[] = {
311 FLAG_INFO(FULL_THREAD
),
313 FLAG_INFO(FIRST_ATEST
),
319 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
320 static const struct pandecode_flag_info mfbd_flag_info
[] = {
321 FLAG_INFO(DEPTH_WRITE
),
327 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
328 static const struct pandecode_flag_info sfbd_unk1_info
[] = {
335 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
336 static const struct pandecode_flag_info sfbd_unk2_info
[] = {
343 /* Midgard's tiler descriptor is embedded within the
347 pandecode_midgard_tiler_descriptor(
348 const struct midgard_tiler_descriptor
*t
,
354 pandecode_log(".tiler = {\n");
357 if (t
->hierarchy_mask
== MALI_TILER_DISABLED
)
358 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
360 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
362 /* We know this name from the kernel, but we never see it nonzero */
365 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16
, t
->flags
);
367 MEMORY_PROP(t
, polygon_list
);
369 /* The body is offset from the base of the polygon list */
370 //assert(t->polygon_list_body > t->polygon_list);
371 unsigned body_offset
= t
->polygon_list_body
- t
->polygon_list
;
373 /* It needs to fit inside the reported size */
374 //assert(t->polygon_list_size >= body_offset);
376 /* Now that we've sanity checked, we'll try to calculate the sizes
377 * ourselves for comparison */
379 unsigned ref_header
= panfrost_tiler_header_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
380 unsigned ref_size
= panfrost_tiler_full_size(width
, height
, t
->hierarchy_mask
, has_hierarchy
);
382 if (!((ref_header
== body_offset
) && (ref_size
== t
->polygon_list_size
))) {
383 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
384 ref_header
, ref_size
);
385 pandecode_prop("polygon_list_size = 0x%x", t
->polygon_list_size
);
386 pandecode_msg("body offset %d\n", body_offset
);
389 /* The tiler heap has a start and end specified -- it should be
390 * identical to what we have in the BO. The exception is if tiling is
393 MEMORY_PROP(t
, heap_start
);
394 assert(t
->heap_end
>= t
->heap_start
);
396 struct pandecode_mapped_memory
*heap
=
397 pandecode_find_mapped_gpu_mem_containing(t
->heap_start
);
399 unsigned heap_size
= t
->heap_end
- t
->heap_start
;
401 /* Tiling is enabled with a special flag */
402 unsigned hierarchy_mask
= t
->hierarchy_mask
& MALI_HIERARCHY_MASK
;
403 unsigned tiler_flags
= t
->hierarchy_mask
^ hierarchy_mask
;
405 bool tiling_enabled
= hierarchy_mask
;
407 if (tiling_enabled
) {
408 /* When tiling is enabled, the heap should be a tight fit */
409 unsigned heap_offset
= t
->heap_start
- heap
->gpu_va
;
410 if ((heap_offset
+ heap_size
) != heap
->length
) {
411 pandecode_msg("XXX: heap size %u (expected %zu)\n",
412 heap_size
, heap
->length
- heap_offset
);
415 /* We should also have no other flags */
417 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags
);
419 /* When tiling is disabled, we should have that flag and no others */
421 if (tiler_flags
!= MALI_TILER_DISABLED
) {
422 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
426 /* We should also have an empty heap */
428 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
432 /* Disabled tiling is used only for clear-only jobs, which are
433 * purely FRAGMENT, so we should never see this for
434 * non-FRAGMENT descriptors. */
437 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
440 /* We've never seen weights used in practice, but we know from the
441 * kernel these fields is there */
443 bool nonzero_weights
= false;
445 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
446 nonzero_weights
|= t
->weights
[w
] != 0x0;
449 if (nonzero_weights
) {
450 pandecode_log(".weights = { ");
452 for (unsigned w
= 0; w
< ARRAY_SIZE(t
->weights
); ++w
) {
453 pandecode_log_cont("%d, ", t
->weights
[w
]);
460 pandecode_log("}\n");
463 /* TODO: The Bifrost tiler is not understood at all yet */
466 pandecode_bifrost_tiler_descriptor(const struct mali_framebuffer
*fb
)
468 pandecode_log(".tiler = {\n");
471 MEMORY_PROP(fb
, tiler_meta
);
473 for (int i
= 0; i
< 16; i
++) {
474 if (fb
->zeros
[i
] != 0) {
475 pandecode_msg("XXX: tiler descriptor zero %d tripped, value %x\n",
480 pandecode_log("},\n");
483 pandecode_log("}\n");
487 /* Information about the framebuffer passed back for
488 * additional analysis */
490 struct pandecode_fbd
{
498 pandecode_sfbd_format(struct mali_sfbd_format format
)
500 pandecode_log(".format = {\n");
503 pandecode_log(".unk1 = ");
504 pandecode_log_decoded_flags(sfbd_unk1_info
, format
.unk1
);
505 pandecode_log_cont(",\n");
507 /* TODO: Map formats so we can check swizzles and print nicely */
508 pandecode_log("swizzle");
509 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
510 pandecode_log_cont(",\n");
512 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
513 (format
.nr_channels
+ 1));
515 pandecode_log(".unk2 = ");
516 pandecode_log_decoded_flags(sfbd_unk2_info
, format
.unk2
);
517 pandecode_log_cont(",\n");
519 pandecode_prop("block = %s", mali_block_format_as_str(format
.block
));
521 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
524 pandecode_log("},\n");
528 pandecode_shared_memory(const struct mali_shared_memory
*desc
, bool is_compute
)
530 pandecode_prop("stack_shift = 0x%x", desc
->stack_shift
);
533 pandecode_prop("unk0 = 0x%x", desc
->unk0
);
535 if (desc
->shared_workgroup_count
!= 0x1F) {
536 pandecode_prop("shared_workgroup_count = %d", desc
->shared_workgroup_count
);
538 pandecode_msg("XXX: wrong workgroup count for noncompute\n");
541 if (desc
->shared_unk1
|| desc
->shared_shift
) {
542 pandecode_prop("shared_unk1 = %X", desc
->shared_unk1
);
543 pandecode_prop("shared_shift = %X", desc
->shared_shift
);
546 pandecode_msg("XXX: shared memory configured in noncompute shader");
549 if (desc
->shared_zero
) {
550 pandecode_msg("XXX: shared memory zero tripped\n");
551 pandecode_prop("shared_zero = 0x%" PRIx32
, desc
->shared_zero
);
554 if (desc
->shared_memory
&& !is_compute
)
555 pandecode_msg("XXX: shared memory used in noncompute shader\n");
557 MEMORY_PROP(desc
, scratchpad
);
558 MEMORY_PROP(desc
, shared_memory
);
559 MEMORY_PROP(desc
, unknown1
);
561 if (desc
->scratchpad
) {
562 struct pandecode_mapped_memory
*smem
=
563 pandecode_find_mapped_gpu_mem_containing(desc
->scratchpad
);
565 pandecode_msg("scratchpad size %u\n", smem
->length
);
570 static struct pandecode_fbd
571 pandecode_sfbd(uint64_t gpu_va
, int job_no
, bool is_fragment
, unsigned gpu_id
)
573 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
574 const struct mali_single_framebuffer
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
576 struct pandecode_fbd info
= {
581 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
584 pandecode_log(".shared_memory = {\n");
586 pandecode_shared_memory(&s
->shared_memory
, false);
588 pandecode_log("},\n");
590 pandecode_sfbd_format(s
->format
);
592 info
.width
= s
->width
+ 1;
593 info
.height
= s
->height
+ 1;
595 pandecode_prop("width = MALI_POSITIVE(%" PRId16
")", info
.width
);
596 pandecode_prop("height = MALI_POSITIVE(%" PRId16
")", info
.height
);
598 MEMORY_PROP(s
, checksum
);
600 if (s
->checksum_stride
)
601 pandecode_prop("checksum_stride = %d", s
->checksum_stride
);
603 MEMORY_PROP(s
, framebuffer
);
604 pandecode_prop("stride = %d", s
->stride
);
606 /* Earlier in the actual commandstream -- right before width -- but we
607 * delay to flow nicer */
609 pandecode_log(".clear_flags = ");
610 pandecode_log_decoded_flags(clear_flag_info
, s
->clear_flags
);
611 pandecode_log_cont(",\n");
613 if (s
->depth_buffer
) {
614 MEMORY_PROP(s
, depth_buffer
);
615 pandecode_prop("depth_stride = %d", s
->depth_stride
);
618 if (s
->stencil_buffer
) {
619 MEMORY_PROP(s
, stencil_buffer
);
620 pandecode_prop("stencil_stride = %d", s
->stencil_stride
);
623 if (s
->depth_stride_zero
||
624 s
->stencil_stride_zero
||
625 s
->zero7
|| s
->zero8
) {
626 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
627 pandecode_prop("depth_stride_zero = 0x%x",
628 s
->depth_stride_zero
);
629 pandecode_prop("stencil_stride_zero = 0x%x",
630 s
->stencil_stride_zero
);
631 pandecode_prop("zero7 = 0x%" PRIx32
,
633 pandecode_prop("zero8 = 0x%" PRIx32
,
637 if (s
->clear_color_1
| s
->clear_color_2
| s
->clear_color_3
| s
->clear_color_4
) {
638 pandecode_prop("clear_color_1 = 0x%" PRIx32
, s
->clear_color_1
);
639 pandecode_prop("clear_color_2 = 0x%" PRIx32
, s
->clear_color_2
);
640 pandecode_prop("clear_color_3 = 0x%" PRIx32
, s
->clear_color_3
);
641 pandecode_prop("clear_color_4 = 0x%" PRIx32
, s
->clear_color_4
);
644 if (s
->clear_depth_1
!= 0 || s
->clear_depth_2
!= 0 || s
->clear_depth_3
!= 0 || s
->clear_depth_4
!= 0) {
645 pandecode_prop("clear_depth_1 = %f", s
->clear_depth_1
);
646 pandecode_prop("clear_depth_2 = %f", s
->clear_depth_2
);
647 pandecode_prop("clear_depth_3 = %f", s
->clear_depth_3
);
648 pandecode_prop("clear_depth_4 = %f", s
->clear_depth_4
);
651 if (s
->clear_stencil
) {
652 pandecode_prop("clear_stencil = 0x%x", s
->clear_stencil
);
655 const struct midgard_tiler_descriptor t
= s
->tiler
;
657 bool has_hierarchy
= !(gpu_id
== 0x0720 || gpu_id
== 0x0820 || gpu_id
== 0x0830);
658 pandecode_midgard_tiler_descriptor(&t
, s
->width
+ 1, s
->height
+ 1, is_fragment
, has_hierarchy
);
661 pandecode_log("};\n");
663 pandecode_prop("zero2 = 0x%" PRIx32
, s
->zero2
);
664 pandecode_prop("zero4 = 0x%" PRIx32
, s
->zero4
);
665 pandecode_prop("zero5 = 0x%" PRIx32
, s
->zero5
);
667 pandecode_log_cont(".zero3 = {");
669 for (int i
= 0; i
< sizeof(s
->zero3
) / sizeof(s
->zero3
[0]); ++i
)
670 pandecode_log_cont("%X, ", s
->zero3
[i
]);
672 pandecode_log_cont("},\n");
674 pandecode_log_cont(".zero6 = {");
676 for (int i
= 0; i
< sizeof(s
->zero6
) / sizeof(s
->zero6
[0]); ++i
)
677 pandecode_log_cont("%X, ", s
->zero6
[i
]);
679 pandecode_log_cont("},\n");
685 pandecode_compute_fbd(uint64_t gpu_va
, int job_no
)
687 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
688 const struct mali_shared_memory
*PANDECODE_PTR_VAR(s
, mem
, (mali_ptr
) gpu_va
);
690 pandecode_log("struct mali_shared_memory shared_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
692 pandecode_shared_memory(s
, true);
694 pandecode_log("},\n");
697 /* Extracts the number of components associated with a Mali format */
700 pandecode_format_component_count(enum mali_format fmt
)
702 /* Mask out the format class */
703 unsigned top
= fmt
& 0b11100000;
706 case MALI_FORMAT_SNORM
:
707 case MALI_FORMAT_UINT
:
708 case MALI_FORMAT_UNORM
:
709 case MALI_FORMAT_SINT
:
710 return ((fmt
>> 3) & 3) + 1;
717 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
720 pandecode_access_mask_from_channel_swizzle(unsigned swizzle
)
723 assert(MALI_CHANNEL_R
== 0);
725 for (unsigned c
= 0; c
< 4; ++c
) {
726 enum mali_channel chan
= (swizzle
>> (3*c
)) & 0x7;
728 if (chan
<= MALI_CHANNEL_A
)
735 /* Validates that a (format, swizzle) pair is valid, in the sense that the
736 * swizzle doesn't access any components that are undefined in the format.
737 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
741 pandecode_validate_format_swizzle(enum mali_format fmt
, unsigned swizzle
)
743 unsigned nr_comp
= pandecode_format_component_count(fmt
);
744 unsigned access_mask
= pandecode_access_mask_from_channel_swizzle(swizzle
);
745 unsigned valid_mask
= (1 << nr_comp
) - 1;
746 unsigned invalid_mask
= ~valid_mask
;
748 if (access_mask
& invalid_mask
) {
749 pandecode_msg("XXX: invalid components accessed\n");
753 /* Check for the default non-swizzling swizzle so we can suppress
754 * useless printing for the defaults */
756 unsigned default_swizzles
[4] = {
757 MALI_CHANNEL_R
| (MALI_CHANNEL_0
<< 3) | (MALI_CHANNEL_0
<< 6) | (MALI_CHANNEL_1
<< 9),
758 MALI_CHANNEL_R
| (MALI_CHANNEL_G
<< 3) | (MALI_CHANNEL_0
<< 6) | (MALI_CHANNEL_1
<< 9),
759 MALI_CHANNEL_R
| (MALI_CHANNEL_G
<< 3) | (MALI_CHANNEL_B
<< 6) | (MALI_CHANNEL_1
<< 9),
760 MALI_CHANNEL_R
| (MALI_CHANNEL_G
<< 3) | (MALI_CHANNEL_B
<< 6) | (MALI_CHANNEL_A
<< 9)
763 return (swizzle
== default_swizzles
[nr_comp
- 1]);
767 pandecode_swizzle(unsigned swizzle
, enum mali_format format
)
769 /* First, do some validation */
770 bool trivial_swizzle
= pandecode_validate_format_swizzle(
776 /* Next, print the swizzle */
777 pandecode_log_cont(".");
779 static const char components
[] = "rgba01";
781 for (unsigned c
= 0; c
< 4; ++c
) {
782 enum mali_channel chan
= (swizzle
>> (3 * c
)) & 0x7;
784 if (chan
> MALI_CHANNEL_1
) {
785 pandecode_log("XXX: invalid swizzle channel %d\n", chan
);
788 pandecode_log_cont("%c", components
[chan
]);
793 pandecode_rt_format(struct mali_rt_format format
)
795 pandecode_log(".format = {\n");
798 pandecode_prop("unk1 = 0x%" PRIx32
, format
.unk1
);
799 pandecode_prop("unk2 = 0x%" PRIx32
, format
.unk2
);
800 pandecode_prop("unk3 = 0x%" PRIx32
, format
.unk3
);
801 pandecode_prop("unk4 = 0x%" PRIx32
, format
.unk4
);
803 pandecode_prop("block = %s", mali_block_format_as_str(format
.block
));
805 /* TODO: Map formats so we can check swizzles and print nicely */
806 pandecode_log("swizzle");
807 pandecode_swizzle(format
.swizzle
, MALI_RGBA8_UNORM
);
808 pandecode_log_cont(",\n");
810 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
811 (format
.nr_channels
+ 1));
813 pandecode_log(".flags = ");
814 pandecode_log_decoded_flags(mfbd_fmt_flag_info
, format
.flags
);
815 pandecode_log_cont(",\n");
817 pandecode_prop("msaa = %s", mali_msaa_as_str(format
.msaa
));
819 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
820 * which is a faster hardware-based alternative to the wallpaper method
821 * to preserve framebuffer contents across frames. In practice, MFBD
822 * preload is buggy on Midgard, and so this is a chicken bit. If this
823 * bit isn't set, most likely something broke unrelated to preload */
825 if (!format
.no_preload
) {
826 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
827 pandecode_prop("no_preload = 0x%" PRIx32
, format
.no_preload
);
831 pandecode_prop("zero = 0x%" PRIx32
, format
.zero
);
834 pandecode_log("},\n");
838 pandecode_render_target(uint64_t gpu_va
, unsigned job_no
, const struct mali_framebuffer
*fb
)
840 pandecode_log("struct mali_render_target rts_list_%"PRIx64
"_%d[] = {\n", gpu_va
, job_no
);
843 for (int i
= 0; i
< (fb
->rt_count_1
+ 1); i
++) {
844 mali_ptr rt_va
= gpu_va
+ i
* sizeof(struct mali_render_target
);
845 struct pandecode_mapped_memory
*mem
=
846 pandecode_find_mapped_gpu_mem_containing(rt_va
);
847 const struct mali_render_target
*PANDECODE_PTR_VAR(rt
, mem
, (mali_ptr
) rt_va
);
849 pandecode_log("{\n");
852 pandecode_rt_format(rt
->format
);
854 if (rt
->format
.block
== MALI_BLOCK_FORMAT_AFBC
) {
855 pandecode_log(".afbc = {\n");
858 char *a
= pointer_as_memory_reference(rt
->afbc
.metadata
);
859 pandecode_prop("metadata = %s", a
);
862 pandecode_prop("stride = %d", rt
->afbc
.stride
);
864 pandecode_log(".flags = ");
865 pandecode_log_decoded_flags(afbc_fmt_flag_info
, rt
->afbc
.flags
);
866 pandecode_log_cont(",\n");
869 pandecode_log("},\n");
870 } else if (rt
->afbc
.metadata
|| rt
->afbc
.stride
|| rt
->afbc
.flags
) {
871 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
877 MEMORY_PROP(rt
, framebuffer
);
878 pandecode_prop("framebuffer_stride = %d", rt
->framebuffer_stride
);
880 if (rt
->layer_stride
)
881 pandecode_prop("layer_stride = %d", rt
->layer_stride
);
883 if (rt
->clear_color_1
| rt
->clear_color_2
| rt
->clear_color_3
| rt
->clear_color_4
) {
884 pandecode_prop("clear_color_1 = 0x%" PRIx32
, rt
->clear_color_1
);
885 pandecode_prop("clear_color_2 = 0x%" PRIx32
, rt
->clear_color_2
);
886 pandecode_prop("clear_color_3 = 0x%" PRIx32
, rt
->clear_color_3
);
887 pandecode_prop("clear_color_4 = 0x%" PRIx32
, rt
->clear_color_4
);
890 if (rt
->zero1
|| rt
->zero2
) {
891 pandecode_msg("XXX: render target zeros tripped\n");
892 pandecode_prop("zero1 = 0x%" PRIx64
, rt
->zero1
);
893 pandecode_prop("zero2 = 0x%" PRIx32
, rt
->zero2
);
897 pandecode_log("},\n");
901 pandecode_log("};\n");
904 static struct pandecode_fbd
905 pandecode_mfbd_bfr(uint64_t gpu_va
, int job_no
, bool is_fragment
, bool is_compute
, bool is_bifrost
)
907 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
908 const struct mali_framebuffer
*PANDECODE_PTR_VAR(fb
, mem
, (mali_ptr
) gpu_va
);
910 struct pandecode_fbd info
;
912 if (is_bifrost
&& fb
->msaa
.sample_locations
) {
913 /* The blob stores all possible sample locations in a single buffer
914 * allocated on startup, and just switches the pointer when switching
915 * MSAA state. For now, we just put the data into the cmdstream, but we
916 * should do something like what the blob does with a real driver.
918 * There seem to be 32 slots for sample locations, followed by another
919 * 16. The second 16 is just the center location followed by 15 zeros
920 * in all the cases I've identified (maybe shader vs. depth/color
924 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(fb
->msaa
.sample_locations
);
926 const u16
*PANDECODE_PTR_VAR(samples
, smem
, fb
->msaa
.sample_locations
);
928 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no
);
931 for (int i
= 0; i
< 32 + 16; i
++) {
932 pandecode_log("%d, %d,\n", samples
[2 * i
], samples
[2 * i
+ 1]);
936 pandecode_log("};\n");
939 pandecode_log("struct mali_framebuffer framebuffer_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
943 pandecode_log(".msaa = {\n");
946 if (fb
->msaa
.sample_locations
)
947 pandecode_prop("sample_locations = sample_locations_%d", job_no
);
949 pandecode_msg("XXX: sample_locations missing\n");
951 if (fb
->msaa
.zero1
|| fb
->msaa
.zero2
|| fb
->msaa
.zero4
) {
952 pandecode_msg("XXX: multisampling zero tripped\n");
953 pandecode_prop("zero1 = %" PRIx64
, fb
->msaa
.zero1
);
954 pandecode_prop("zero2 = %" PRIx64
, fb
->msaa
.zero2
);
955 pandecode_prop("zero4 = %" PRIx64
, fb
->msaa
.zero4
);
959 pandecode_log("},\n");
961 pandecode_log(".shared_memory = {\n");
963 pandecode_shared_memory(&fb
->shared_memory
, is_compute
);
965 pandecode_log("},\n");
968 info
.width
= fb
->width1
+ 1;
969 info
.height
= fb
->height1
+ 1;
970 info
.rt_count
= fb
->rt_count_1
+ 1;
972 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb
->width1
+ 1);
973 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb
->height1
+ 1);
974 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb
->width2
+ 1);
975 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb
->height2
+ 1);
977 pandecode_prop("unk1 = 0x%x", fb
->unk1
);
978 pandecode_prop("unk2 = 0x%x", fb
->unk2
);
979 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb
->rt_count_1
+ 1);
980 pandecode_prop("rt_count_2 = %d", fb
->rt_count_2
);
982 pandecode_log(".mfbd_flags = ");
983 pandecode_log_decoded_flags(mfbd_flag_info
, fb
->mfbd_flags
);
984 pandecode_log_cont(",\n");
986 if (fb
->clear_stencil
)
987 pandecode_prop("clear_stencil = 0x%x", fb
->clear_stencil
);
990 pandecode_prop("clear_depth = %f", fb
->clear_depth
);
994 pandecode_bifrost_tiler_descriptor(fb
);
996 const struct midgard_tiler_descriptor t
= fb
->tiler
;
997 pandecode_midgard_tiler_descriptor(&t
, fb
->width1
+ 1, fb
->height1
+ 1, is_fragment
, true);
1000 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
1002 if (fb
->zero3
|| fb
->zero4
) {
1003 pandecode_msg("XXX: framebuffer zeros tripped\n");
1004 pandecode_prop("zero3 = 0x%" PRIx32
, fb
->zero3
);
1005 pandecode_prop("zero4 = 0x%" PRIx32
, fb
->zero4
);
1009 pandecode_log("};\n");
1011 gpu_va
+= sizeof(struct mali_framebuffer
);
1013 info
.has_extra
= (fb
->mfbd_flags
& MALI_MFBD_EXTRA
) && is_fragment
;
1015 if (info
.has_extra
) {
1016 mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
1017 const struct mali_framebuffer_extra
*PANDECODE_PTR_VAR(fbx
, mem
, (mali_ptr
) gpu_va
);
1019 pandecode_log("struct mali_framebuffer_extra fb_extra_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
1022 MEMORY_PROP(fbx
, checksum
);
1024 if (fbx
->checksum_stride
)
1025 pandecode_prop("checksum_stride = %d", fbx
->checksum_stride
);
1027 pandecode_log(".flags_hi = ");
1028 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info
, fbx
->flags_hi
);
1029 pandecode_log_cont(",\n");
1031 pandecode_log(".flags_lo = ");
1032 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info
, fbx
->flags_lo
);
1033 pandecode_log_cont(",\n");
1035 pandecode_prop("zs_block = %s", mali_block_format_as_str(fbx
->zs_block
));
1036 pandecode_prop("zs_samples = MALI_POSITIVE(%u)", fbx
->zs_samples
+ 1);
1038 if (fbx
->zs_block
== MALI_BLOCK_FORMAT_AFBC
) {
1039 pandecode_log(".ds_afbc = {\n");
1042 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil_afbc_metadata
);
1043 pandecode_prop("depth_stencil_afbc_stride = %d",
1044 fbx
->ds_afbc
.depth_stencil_afbc_stride
);
1045 MEMORY_PROP_DIR(fbx
->ds_afbc
, depth_stencil
);
1047 pandecode_log(".flags = ");
1048 pandecode_log_decoded_flags(afbc_fmt_flag_info
, fbx
->ds_afbc
.flags
);
1049 pandecode_log_cont(",\n");
1051 if (fbx
->ds_afbc
.padding
) {
1052 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1053 pandecode_prop("padding = 0x%" PRIx64
, fbx
->ds_afbc
.padding
);
1057 pandecode_log("},\n");
1059 pandecode_log(".ds_linear = {\n");
1062 if (fbx
->ds_linear
.depth
) {
1063 MEMORY_PROP_DIR(fbx
->ds_linear
, depth
);
1064 pandecode_prop("depth_stride = %d",
1065 fbx
->ds_linear
.depth_stride
);
1066 pandecode_prop("depth_layer_stride = %d",
1067 fbx
->ds_linear
.depth_layer_stride
);
1068 } else if (fbx
->ds_linear
.depth_stride
|| fbx
->ds_linear
.depth_layer_stride
) {
1069 pandecode_msg("XXX: depth stride zero tripped %d %d\n", fbx
->ds_linear
.depth_stride
, fbx
->ds_linear
.depth_layer_stride
);
1072 if (fbx
->ds_linear
.stencil
) {
1073 MEMORY_PROP_DIR(fbx
->ds_linear
, stencil
);
1074 pandecode_prop("stencil_stride = %d",
1075 fbx
->ds_linear
.stencil_stride
);
1076 pandecode_prop("stencil_layer_stride = %d",
1077 fbx
->ds_linear
.stencil_layer_stride
);
1078 } else if (fbx
->ds_linear
.stencil_stride
|| fbx
->ds_linear
.stencil_layer_stride
) {
1079 pandecode_msg("XXX: stencil stride zero tripped %d %d\n", fbx
->ds_linear
.stencil_stride
, fbx
->ds_linear
.stencil_layer_stride
);
1082 if (fbx
->ds_linear
.depth_stride_zero
||
1083 fbx
->ds_linear
.stencil_stride_zero
) {
1084 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1085 pandecode_prop("depth_stride_zero = 0x%x",
1086 fbx
->ds_linear
.depth_stride_zero
);
1087 pandecode_prop("stencil_stride_zero = 0x%x",
1088 fbx
->ds_linear
.stencil_stride_zero
);
1092 pandecode_log("},\n");
1095 if (fbx
->clear_color_1
| fbx
->clear_color_2
) {
1096 pandecode_prop("clear_color_1 = 0x%" PRIx32
, fbx
->clear_color_1
);
1097 pandecode_prop("clear_color_2 = 0x%" PRIx32
, fbx
->clear_color_2
);
1101 pandecode_msg("XXX: fb_extra zeros tripped\n");
1102 pandecode_prop("zero3 = 0x%" PRIx64
, fbx
->zero3
);
1106 pandecode_log("};\n");
1108 gpu_va
+= sizeof(struct mali_framebuffer_extra
);
1112 pandecode_render_target(gpu_va
, job_no
, fb
);
1118 pandecode_attributes(const struct pandecode_mapped_memory
*mem
,
1119 mali_ptr addr
, int job_no
, char *suffix
,
1120 int count
, bool varying
, enum mali_job_type job_type
)
1122 char *prefix
= varying
? "Varying" : "Attribute";
1126 pandecode_msg("warn: No %s records\n", prefix
);
1130 MAP_ADDR(ATTRIBUTE_BUFFER
, addr
, cl
);
1132 for (int i
= 0; i
< count
; ++i
) {
1133 fprintf(pandecode_dump_stream
, "%s\n", prefix
);
1135 struct MALI_ATTRIBUTE_BUFFER temp
;
1136 MALI_ATTRIBUTE_BUFFER_unpack(cl
+ i
* MALI_ATTRIBUTE_BUFFER_LENGTH
, &temp
);
1137 MALI_ATTRIBUTE_BUFFER_print(pandecode_dump_stream
, &temp
, 2);
1139 if (temp
.type
== MALI_ATTRIBUTE_TYPE_1D_NPOT_DIVISOR
) {
1140 struct MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT temp2
;
1141 MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT_unpack(cl
+ (i
+ 1) * MALI_ATTRIBUTE_BUFFER_LENGTH
, &temp2
);
1142 MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT_print(pandecode_dump_stream
, &temp2
, 2);
1148 pandecode_shader_address(const char *name
, mali_ptr ptr
)
1150 /* TODO: Decode flags */
1151 mali_ptr shader_ptr
= ptr
& ~15;
1153 char *a
= pointer_as_memory_reference(shader_ptr
);
1154 pandecode_prop("%s = (%s) | %d", name
, a
, (int) (ptr
& 15));
1161 pandecode_blend_equation(const struct mali_blend_equation
*blend
)
1164 pandecode_msg("XXX: blend zero tripped: %X\n", blend
->zero1
);
1166 pandecode_log(".equation = {\n");
1169 pandecode_prop("rgb_mode = 0x%X", blend
->rgb_mode
);
1170 pandecode_prop("alpha_mode = 0x%X", blend
->alpha_mode
);
1172 pandecode_log(".color_mask = ");
1173 pandecode_log_decoded_flags(mask_flag_info
, blend
->color_mask
);
1174 pandecode_log_cont(",\n");
1177 pandecode_log("},\n");
1180 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1183 decode_bifrost_constant(u16 constant
)
1185 float lo
= (float) (constant
& 0xFF);
1186 float hi
= (float) (constant
>> 8);
1188 return (hi
/ 255.0) + (lo
/ 65535.0);
1192 pandecode_bifrost_blend(void *descs
, int job_no
, int rt_no
)
1194 struct bifrost_blend_rt
*b
=
1195 ((struct bifrost_blend_rt
*) descs
) + rt_no
;
1197 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1200 pandecode_prop("flags = 0x%" PRIx16
, b
->flags
);
1201 pandecode_prop("constant = 0x%" PRIx8
" /* %f */",
1202 b
->constant
, decode_bifrost_constant(b
->constant
));
1204 /* TODO figure out blend shader enable bit */
1205 pandecode_blend_equation(&b
->equation
);
1207 pandecode_prop("unk2 = 0x%" PRIx16
, b
->unk2
);
1208 pandecode_prop("index = 0x%" PRIx16
, b
->index
);
1210 pandecode_log(".format = %s", mali_format_as_str(b
->format
));
1211 pandecode_swizzle(b
->swizzle
, b
->format
);
1212 pandecode_log_cont(",\n");
1214 pandecode_prop("swizzle = 0x%" PRIx32
, b
->swizzle
);
1215 pandecode_prop("format = 0x%" PRIx32
, b
->format
);
1218 pandecode_msg("XXX: pandecode_bifrost_blend zero1 tripped\n");
1219 pandecode_prop("zero1 = 0x%" PRIx32
, b
->zero1
);
1222 pandecode_log(".shader_type = ");
1223 switch(b
->shader_type
) {
1224 case BIFROST_BLEND_F16
:
1225 pandecode_log_cont("BIFROST_BLEND_F16");
1227 case BIFROST_BLEND_F32
:
1228 pandecode_log_cont("BIFROST_BLEND_F32");
1230 case BIFROST_BLEND_I32
:
1231 pandecode_log_cont("BIFROST_BLEND_I32");
1233 case BIFROST_BLEND_U32
:
1234 pandecode_log_cont("BIFROST_BLEND_U32");
1236 case BIFROST_BLEND_I16
:
1237 pandecode_log_cont("BIFROST_BLEND_I16");
1239 case BIFROST_BLEND_U16
:
1240 pandecode_log_cont("BIFROST_BLEND_U16");
1243 pandecode_log_cont(",\n");
1246 pandecode_msg("XXX: pandecode_bifrost_blend zero2 tripped\n");
1247 pandecode_prop("zero2 = 0x%" PRIx32
, b
->zero2
);
1250 pandecode_prop("shader = 0x%" PRIx32
, b
->shader
);
1253 pandecode_log("},\n");
1259 pandecode_midgard_blend(union midgard_blend
*blend
, bool is_shader
)
1261 /* constant/equation is in a union */
1265 pandecode_log(".blend = {\n");
1269 pandecode_shader_address("shader", blend
->shader
);
1271 pandecode_blend_equation(&blend
->equation
);
1272 pandecode_prop("constant = %f", blend
->constant
);
1276 pandecode_log("},\n");
1278 /* Return blend shader to disassemble if present */
1279 return is_shader
? (blend
->shader
& ~0xF) : 0;
1283 pandecode_midgard_blend_mrt(void *descs
, int job_no
, int rt_no
)
1285 struct midgard_blend_rt
*b
=
1286 ((struct midgard_blend_rt
*) descs
) + rt_no
;
1288 /* Flags determine presence of blend shader */
1289 bool is_shader
= (b
->flags
& 0xF) >= 0x2;
1291 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no
, rt_no
);
1294 pandecode_prop("flags = 0x%" PRIx64
, b
->flags
);
1296 union midgard_blend blend
= b
->blend
;
1297 mali_ptr shader
= pandecode_midgard_blend(&blend
, is_shader
);
1300 pandecode_log("};\n");
1305 /* Attributes and varyings have descriptor records, which contain information
1306 * about their format and ordering with the attribute/varying buffers. We'll
1307 * want to validate that the combinations specified are self-consistent.
1311 pandecode_attribute_meta(int job_no
, int count
, const struct mali_vertex_tiler_postfix
*v
, bool varying
, char *suffix
)
1313 const char *prefix
= varying
? "Varying" : "Attribute";
1314 mali_ptr p
= varying
? v
->varying_meta
: v
->attribute_meta
;
1316 for (int i
= 0; i
< count
; ++i
, p
+= MALI_ATTRIBUTE_LENGTH
)
1317 DUMP_ADDR(prefix
, ATTRIBUTE
, p
, 1);
1322 /* return bits [lo, hi) of word */
1324 bits(u32 word
, u32 lo
, u32 hi
)
1327 return word
; // avoid undefined behavior with the shift
1329 return (word
>> lo
) & ((1 << (hi
- lo
)) - 1);
1333 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix
*p
, int job_no
, bool graphics
)
1335 pandecode_log(".prefix = {\n");
1338 /* Decode invocation_count. See the comment before the definition of
1339 * invocation_count for an explanation.
1342 unsigned size_y_shift
= bits(p
->invocation_shifts
, 0, 5);
1343 unsigned size_z_shift
= bits(p
->invocation_shifts
, 5, 10);
1344 unsigned workgroups_x_shift
= bits(p
->invocation_shifts
, 10, 16);
1345 unsigned workgroups_y_shift
= bits(p
->invocation_shifts
, 16, 22);
1346 unsigned workgroups_z_shift
= bits(p
->invocation_shifts
, 22, 28);
1347 unsigned workgroups_x_shift_2
= bits(p
->invocation_shifts
, 28, 32);
1349 unsigned size_x
= bits(p
->invocation_count
, 0, size_y_shift
) + 1;
1350 unsigned size_y
= bits(p
->invocation_count
, size_y_shift
, size_z_shift
) + 1;
1351 unsigned size_z
= bits(p
->invocation_count
, size_z_shift
, workgroups_x_shift
) + 1;
1353 unsigned groups_x
= bits(p
->invocation_count
, workgroups_x_shift
, workgroups_y_shift
) + 1;
1354 unsigned groups_y
= bits(p
->invocation_count
, workgroups_y_shift
, workgroups_z_shift
) + 1;
1355 unsigned groups_z
= bits(p
->invocation_count
, workgroups_z_shift
, 32) + 1;
1357 /* Even though we have this decoded, we want to ensure that the
1358 * representation is "unique" so we don't lose anything by printing only
1359 * the final result. More specifically, we need to check that we were
1360 * passed something in canonical form, since the definition per the
1361 * hardware is inherently not unique. How? Well, take the resulting
1362 * decode and pack it ourselves! If it is bit exact with what we
1363 * decoded, we're good to go. */
1365 struct mali_vertex_tiler_prefix ref
;
1366 panfrost_pack_work_groups_compute(&ref
, groups_x
, groups_y
, groups_z
, size_x
, size_y
, size_z
, graphics
);
1369 (p
->invocation_count
== ref
.invocation_count
) &&
1370 (p
->invocation_shifts
== ref
.invocation_shifts
);
1373 pandecode_msg("XXX: non-canonical workgroups packing\n");
1374 pandecode_msg("expected: %X, %X",
1375 ref
.invocation_count
,
1376 ref
.invocation_shifts
);
1378 pandecode_prop("invocation_count = 0x%" PRIx32
, p
->invocation_count
);
1379 pandecode_prop("size_y_shift = %d", size_y_shift
);
1380 pandecode_prop("size_z_shift = %d", size_z_shift
);
1381 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift
);
1382 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift
);
1383 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift
);
1384 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2
);
1387 /* Regardless, print the decode */
1388 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1389 size_x
, size_y
, size_z
,
1390 groups_x
, groups_y
, groups_z
);
1393 if (p
->unknown_draw
)
1394 pandecode_prop("unknown_draw = 0x%" PRIx32
, p
->unknown_draw
);
1396 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32
, p
->workgroups_x_shift_3
);
1398 if (p
->draw_mode
!= MALI_DRAW_MODE_NONE
)
1399 pandecode_prop("draw_mode = %s", mali_draw_mode_as_str(p
->draw_mode
));
1401 /* Index count only exists for tiler jobs anyway */
1404 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32
")", p
->index_count
+ 1);
1407 unsigned index_raw_size
= (p
->unknown_draw
& MALI_DRAW_INDEXED_SIZE
);
1408 index_raw_size
>>= MALI_DRAW_INDEXED_SHIFT
;
1410 /* Validate an index buffer is present if we need one. TODO: verify
1411 * relationship between invocation_count and index_count */
1414 unsigned count
= p
->index_count
;
1417 unsigned size
= (index_raw_size
== 0x3) ? 4 : index_raw_size
;
1419 /* Ensure we got a size, and if so, validate the index buffer
1420 * is large enough to hold a full set of indices of the given
1423 if (!index_raw_size
)
1424 pandecode_msg("XXX: index size missing\n");
1426 pandecode_validate_buffer(p
->indices
, count
* size
);
1427 } else if (index_raw_size
)
1428 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size
);
1430 if (p
->offset_bias_correction
)
1431 pandecode_prop("offset_bias_correction = %d", p
->offset_bias_correction
);
1433 /* TODO: Figure out what this is. It's not zero */
1434 pandecode_prop("zero1 = 0x%" PRIx32
, p
->zero1
);
1437 pandecode_log("},\n");
1441 pandecode_uniform_buffers(mali_ptr pubufs
, int ubufs_count
, int job_no
)
1443 struct pandecode_mapped_memory
*umem
= pandecode_find_mapped_gpu_mem_containing(pubufs
);
1444 uint64_t *PANDECODE_PTR_VAR(ubufs
, umem
, pubufs
);
1446 for (int i
= 0; i
< ubufs_count
; i
++) {
1447 unsigned size
= (ubufs
[i
] & ((1 << 10) - 1)) * 16;
1448 mali_ptr addr
= (ubufs
[i
] >> 10) << 2;
1450 pandecode_validate_buffer(addr
, size
);
1452 char *ptr
= pointer_as_memory_reference(addr
);
1453 pandecode_log("ubuf_%d[%u] = %s;\n", i
, size
, ptr
);
1457 pandecode_log("\n");
1461 pandecode_uniforms(mali_ptr uniforms
, unsigned uniform_count
)
1463 pandecode_validate_buffer(uniforms
, uniform_count
* 16);
1465 char *ptr
= pointer_as_memory_reference(uniforms
);
1466 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count
, ptr
);
1471 shader_type_for_job(unsigned type
)
1474 case MALI_JOB_TYPE_VERTEX
: return "VERTEX";
1475 case MALI_JOB_TYPE_TILER
: return "FRAGMENT";
1476 case MALI_JOB_TYPE_COMPUTE
: return "COMPUTE";
1482 static unsigned shader_id
= 0;
1484 static struct midgard_disasm_stats
1485 pandecode_shader_disassemble(mali_ptr shader_ptr
, int shader_no
, int type
,
1486 bool is_bifrost
, unsigned gpu_id
)
1488 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(shader_ptr
);
1489 uint8_t *PANDECODE_PTR_VAR(code
, mem
, shader_ptr
);
1491 /* Compute maximum possible size */
1492 size_t sz
= mem
->length
- (shader_ptr
- mem
->gpu_va
);
1494 /* Print some boilerplate to clearly denote the assembly (which doesn't
1495 * obey indentation rules), and actually do the disassembly! */
1497 pandecode_log_cont("\n\n");
1499 struct midgard_disasm_stats stats
;
1502 disassemble_bifrost(pandecode_dump_stream
, code
, sz
, true);
1504 /* TODO: Extend stats to Bifrost */
1505 stats
.texture_count
= -128;
1506 stats
.sampler_count
= -128;
1507 stats
.attribute_count
= -128;
1508 stats
.varying_count
= -128;
1509 stats
.uniform_count
= -128;
1510 stats
.uniform_buffer_count
= -128;
1511 stats
.work_count
= -128;
1513 stats
.instruction_count
= 0;
1514 stats
.bundle_count
= 0;
1515 stats
.quadword_count
= 0;
1516 stats
.helper_invocations
= false;
1518 stats
= disassemble_midgard(pandecode_dump_stream
,
1520 type
== MALI_JOB_TYPE_TILER
?
1521 MESA_SHADER_FRAGMENT
: MESA_SHADER_VERTEX
);
1524 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
1525 * driver-internal purposes with the blob and interfere */
1527 bool should_shaderdb
= type
!= MALI_JOB_TYPE_COMPUTE
;
1529 if (should_shaderdb
) {
1530 unsigned nr_threads
=
1531 (stats
.work_count
<= 4) ? 4 :
1532 (stats
.work_count
<= 8) ? 2 :
1535 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
1536 "%u inst, %u bundles, %u quadwords, "
1537 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
1539 shader_type_for_job(type
),
1540 stats
.instruction_count
, stats
.bundle_count
, stats
.quadword_count
,
1541 stats
.work_count
, nr_threads
);
1549 pandecode_texture_payload(mali_ptr payload
,
1550 enum mali_texture_dimension dim
,
1551 enum mali_texture_layout layout
,
1555 uint16_t array_size
,
1556 struct pandecode_mapped_memory
*tmem
)
1558 pandecode_log(".payload = {\n");
1561 /* A bunch of bitmap pointers follow.
1562 * We work out the correct number,
1563 * based on the mipmap/cubemap
1564 * properties, but dump extra
1565 * possibilities to futureproof */
1567 int bitmap_count
= levels
+ 1;
1569 /* Miptree for each face */
1570 if (dim
== MALI_TEXTURE_DIMENSION_CUBE
)
1573 /* Array of layers */
1574 bitmap_count
*= depth
;
1576 /* Array of textures */
1577 bitmap_count
*= array_size
;
1579 /* Stride for each element */
1583 mali_ptr
*pointers_and_strides
= pandecode_fetch_gpu_mem(tmem
,
1584 payload
, sizeof(mali_ptr
) * bitmap_count
);
1585 for (int i
= 0; i
< bitmap_count
; ++i
) {
1586 /* How we dump depends if this is a stride or a pointer */
1588 if (manual_stride
&& (i
& 1)) {
1589 /* signed 32-bit snuck in as a 64-bit pointer */
1590 uint64_t stride_set
= pointers_and_strides
[i
];
1591 uint32_t clamped_stride
= stride_set
;
1592 int32_t stride
= clamped_stride
;
1593 assert(stride_set
== clamped_stride
);
1594 pandecode_log("(mali_ptr) %d /* stride */, \n", stride
);
1596 char *a
= pointer_as_memory_reference(pointers_and_strides
[i
]);
1597 pandecode_log("%s, \n", a
);
1603 pandecode_log("},\n");
1607 pandecode_texture(mali_ptr u
,
1608 struct pandecode_mapped_memory
*tmem
,
1609 unsigned job_no
, unsigned tex
)
1611 struct pandecode_mapped_memory
*mapped_mem
= pandecode_find_mapped_gpu_mem_containing(u
);
1612 const uint8_t *cl
= pandecode_fetch_gpu_mem(mapped_mem
, u
, MALI_MIDGARD_TEXTURE_LENGTH
);
1614 struct MALI_MIDGARD_TEXTURE temp
;
1615 MALI_MIDGARD_TEXTURE_unpack(cl
, &temp
);
1616 MALI_MIDGARD_TEXTURE_print(pandecode_dump_stream
, &temp
, 2);
1618 pandecode_texture_payload(u
+ MALI_MIDGARD_TEXTURE_LENGTH
,
1619 temp
.dimension
, temp
.texel_ordering
, temp
.manual_stride
,
1620 temp
.levels
, temp
.depth
, temp
.array_size
, mapped_mem
);
1624 pandecode_bifrost_texture(
1629 struct MALI_BIFROST_TEXTURE temp
;
1630 MALI_BIFROST_TEXTURE_unpack(cl
, &temp
);
1631 MALI_BIFROST_TEXTURE_print(pandecode_dump_stream
, &temp
, 2);
1633 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(temp
.surfaces
);
1634 pandecode_texture_payload(temp
.surfaces
, temp
.dimension
, temp
.texel_ordering
,
1635 true, temp
.levels
, 1, 1, tmem
);
1638 /* For shader properties like texture_count, we have a claimed property in the shader_meta, and the actual Truth from static analysis (this may just be an upper limit). We validate accordingly */
1641 pandecode_shader_prop(const char *name
, unsigned claim
, signed truth
, bool fuzzy
)
1647 if (fuzzy
&& (truth
< 0))
1648 pandecode_msg("XXX: fuzzy %s, claimed %d, expected %d\n", name
, claim
, truth
);
1650 if ((truth
>= 0) && !fuzzy
) {
1651 pandecode_msg("%s: expected %s = %d, claimed %u\n",
1652 (truth
< claim
) ? "warn" : "XXX",
1653 name
, truth
, claim
);
1654 } else if ((claim
> -truth
) && !fuzzy
) {
1655 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
1656 name
, -truth
, claim
);
1657 } else if (fuzzy
&& (claim
< truth
))
1658 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
1659 name
, truth
, claim
);
1661 pandecode_log(".%s = %" PRId16
, name
, claim
);
1664 pandecode_log_cont(" /* %u used */", truth
);
1666 pandecode_log_cont(",\n");
1670 pandecode_blend_shader_disassemble(mali_ptr shader
, int job_no
, int job_type
,
1671 bool is_bifrost
, unsigned gpu_id
)
1673 struct midgard_disasm_stats stats
=
1674 pandecode_shader_disassemble(shader
, job_no
, job_type
, is_bifrost
, gpu_id
);
1676 bool has_texture
= (stats
.texture_count
> 0);
1677 bool has_sampler
= (stats
.sampler_count
> 0);
1678 bool has_attribute
= (stats
.attribute_count
> 0);
1679 bool has_varying
= (stats
.varying_count
> 0);
1680 bool has_uniform
= (stats
.uniform_count
> 0);
1681 bool has_ubo
= (stats
.uniform_buffer_count
> 0);
1683 if (has_texture
|| has_sampler
)
1684 pandecode_msg("XXX: blend shader accessing textures\n");
1686 if (has_attribute
|| has_varying
)
1687 pandecode_msg("XXX: blend shader accessing interstage\n");
1689 if (has_uniform
|| has_ubo
)
1690 pandecode_msg("XXX: blend shader accessing uniforms\n");
1694 pandecode_textures(mali_ptr textures
, unsigned texture_count
, int job_no
, bool is_bifrost
)
1696 struct pandecode_mapped_memory
*mmem
= pandecode_find_mapped_gpu_mem_containing(textures
);
1701 pandecode_log("Textures (%"PRIx64
"):\n", textures
);
1704 const void *cl
= pandecode_fetch_gpu_mem(mmem
,
1705 textures
, MALI_BIFROST_TEXTURE_LENGTH
*
1708 for (unsigned tex
= 0; tex
< texture_count
; ++tex
) {
1709 pandecode_bifrost_texture(cl
+
1710 MALI_BIFROST_TEXTURE_LENGTH
* tex
,
1714 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
);
1716 for (int tex
= 0; tex
< texture_count
; ++tex
) {
1717 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
+ tex
* sizeof(mali_ptr
));
1718 char *a
= pointer_as_memory_reference(*u
);
1719 pandecode_log("%s,\n", a
);
1723 /* Now, finally, descend down into the texture descriptor */
1724 for (unsigned tex
= 0; tex
< texture_count
; ++tex
) {
1725 mali_ptr
*PANDECODE_PTR_VAR(u
, mmem
, textures
+ tex
* sizeof(mali_ptr
));
1726 struct pandecode_mapped_memory
*tmem
= pandecode_find_mapped_gpu_mem_containing(*u
);
1728 pandecode_texture(*u
, tmem
, job_no
, tex
);
1734 pandecode_samplers(mali_ptr samplers
, unsigned sampler_count
, int job_no
, bool is_bifrost
)
1736 for (int i
= 0; i
< sampler_count
; ++i
) {
1738 DUMP_ADDR("Sampler", BIFROST_SAMPLER
, samplers
+ (MALI_BIFROST_SAMPLER_LENGTH
* i
), 1);
1740 DUMP_ADDR("Sampler", MIDGARD_SAMPLER
, samplers
+ (MALI_MIDGARD_SAMPLER_LENGTH
* i
), 1);
1746 pandecode_vertex_tiler_postfix_pre(
1747 const struct mali_vertex_tiler_postfix
*p
,
1748 int job_no
, enum mali_job_type job_type
,
1749 char *suffix
, bool is_bifrost
, unsigned gpu_id
)
1751 struct pandecode_mapped_memory
*attr_mem
;
1753 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
1754 * are the only things actually needed from the FBD, vertex/tiler jobs
1755 * no longer reference the FBD -- instead, this field points to some
1756 * info about the scratchpad.
1759 struct pandecode_fbd fbd_info
= {
1760 /* Default for Bifrost */
1765 pandecode_log_cont("\t/* %X %/\n", p
->shared_memory
& 1);
1766 pandecode_compute_fbd(p
->shared_memory
& ~1, job_no
);
1767 } else if (p
->shared_memory
& MALI_MFBD
)
1768 fbd_info
= pandecode_mfbd_bfr((u64
) ((uintptr_t) p
->shared_memory
) & FBD_MASK
, job_no
, false, job_type
== MALI_JOB_TYPE_COMPUTE
, false);
1769 else if (job_type
== MALI_JOB_TYPE_COMPUTE
)
1770 pandecode_compute_fbd((u64
) (uintptr_t) p
->shared_memory
, job_no
);
1772 fbd_info
= pandecode_sfbd((u64
) (uintptr_t) p
->shared_memory
, job_no
, false, gpu_id
);
1774 int varying_count
= 0, attribute_count
= 0, uniform_count
= 0, uniform_buffer_count
= 0;
1775 int texture_count
= 0, sampler_count
= 0;
1778 struct pandecode_mapped_memory
*smem
= pandecode_find_mapped_gpu_mem_containing(p
->shader
);
1779 struct mali_shader_meta
*PANDECODE_PTR_VAR(s
, smem
, p
->shader
);
1781 /* Disassemble ahead-of-time to get stats. Initialize with
1782 * stats for the missing-shader case so we get validation
1785 struct midgard_disasm_stats info
= {
1788 .attribute_count
= 0,
1792 .uniform_count
= -128,
1793 .uniform_buffer_count
= 0
1796 if (s
->shader
& ~0xF)
1797 info
= pandecode_shader_disassemble(s
->shader
& ~0xF, job_no
, job_type
, is_bifrost
, gpu_id
);
1799 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64
"_%d%s = {\n", p
->shader
, job_no
, suffix
);
1802 /* Save for dumps */
1803 attribute_count
= s
->attribute_count
;
1804 varying_count
= s
->varying_count
;
1805 texture_count
= s
->texture_count
;
1806 sampler_count
= s
->sampler_count
;
1809 uniform_count
= s
->bifrost2
.uniform_count
;
1810 uniform_buffer_count
= s
->bifrost1
.uniform_buffer_count
;
1812 uniform_count
= s
->midgard1
.uniform_count
;
1813 uniform_buffer_count
= s
->midgard1
.uniform_buffer_count
;
1816 pandecode_shader_address("shader", s
->shader
);
1818 pandecode_shader_prop("texture_count", s
->texture_count
, info
.texture_count
, false);
1819 pandecode_shader_prop("sampler_count", s
->sampler_count
, info
.sampler_count
, false);
1820 pandecode_shader_prop("attribute_count", s
->attribute_count
, info
.attribute_count
, false);
1821 pandecode_shader_prop("varying_count", s
->varying_count
, info
.varying_count
, false);
1822 pandecode_shader_prop("uniform_buffer_count",
1823 uniform_buffer_count
,
1824 info
.uniform_buffer_count
, true);
1827 pandecode_shader_prop("uniform_count",
1829 info
.uniform_count
, false);
1831 pandecode_shader_prop("work_count",
1832 s
->midgard1
.work_count
, info
.work_count
, false);
1836 pandecode_log("bifrost1.unk1 = ");
1837 pandecode_log_decoded_flags(shader_bifrost_info
, s
->bifrost1
.unk1
);
1838 pandecode_log_cont(",\n");
1840 bool helpers
= s
->midgard1
.flags_lo
& MALI_HELPER_INVOCATIONS
;
1842 if (helpers
!= info
.helper_invocations
) {
1843 pandecode_msg("XXX: expected helpers %u but got %u\n",
1844 info
.helper_invocations
, helpers
);
1847 pandecode_log(".midgard1.flags_lo = ");
1848 pandecode_log_decoded_flags(shader_midgard1_flag_lo_info
,
1849 s
->midgard1
.flags_lo
& ~MALI_HELPER_INVOCATIONS
);
1850 pandecode_log_cont(",\n");
1852 pandecode_log(".midgard1.flags_hi = ");
1853 pandecode_log_decoded_flags(shader_midgard1_flag_hi_info
, s
->midgard1
.flags_hi
);
1854 pandecode_log_cont(",\n");
1857 if (s
->depth_units
|| s
->depth_factor
) {
1858 pandecode_prop("depth_factor = %f", s
->depth_factor
);
1859 pandecode_prop("depth_units = %f", s
->depth_units
);
1862 if (s
->coverage_mask
)
1863 pandecode_prop("coverage_mask = 0x%X", s
->coverage_mask
);
1866 pandecode_prop(".unknown2_2 = %X", s
->unknown2_2
);
1868 if (s
->unknown2_3
|| s
->unknown2_4
) {
1869 pandecode_log(".unknown2_3 = ");
1871 int unknown2_3
= s
->unknown2_3
;
1872 int unknown2_4
= s
->unknown2_4
;
1874 /* We're not quite sure what these flags mean without the depth test, if anything */
1876 if (unknown2_3
& (MALI_DEPTH_WRITEMASK
| MALI_DEPTH_FUNC_MASK
)) {
1877 const char *func
= mali_func_as_str(MALI_GET_DEPTH_FUNC(unknown2_3
));
1878 unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
1880 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func
);
1883 pandecode_log_decoded_flags(u3_flag_info
, unknown2_3
);
1884 pandecode_log_cont(",\n");
1886 pandecode_log(".unknown2_4 = ");
1887 pandecode_log_decoded_flags(u4_flag_info
, unknown2_4
);
1888 pandecode_log_cont(",\n");
1891 if (s
->stencil_mask_front
|| s
->stencil_mask_back
) {
1892 pandecode_prop("stencil_mask_front = 0x%02X", s
->stencil_mask_front
);
1893 pandecode_prop("stencil_mask_back = 0x%02X", s
->stencil_mask_back
);
1896 DUMP_CL("Stencil front", STENCIL
, &s
->stencil_front
, 1);
1897 DUMP_CL("Stencil back", STENCIL
, &s
->stencil_back
, 1);
1900 pandecode_log(".bifrost2 = {\n");
1903 pandecode_prop("unk3 = 0x%" PRIx32
, s
->bifrost2
.unk3
);
1904 pandecode_prop("preload_regs = 0x%" PRIx32
, s
->bifrost2
.preload_regs
);
1905 pandecode_prop("uniform_count = %" PRId32
, s
->bifrost2
.uniform_count
);
1906 pandecode_prop("unk4 = 0x%" PRIx32
, s
->bifrost2
.unk4
);
1909 pandecode_log("},\n");
1910 } else if (s
->midgard2
.unknown2_7
) {
1911 pandecode_log(".midgard2 = {\n");
1914 pandecode_prop("unknown2_7 = 0x%" PRIx32
, s
->midgard2
.unknown2_7
);
1916 pandecode_log("},\n");
1920 pandecode_msg("XXX: shader padding tripped\n");
1921 pandecode_prop("padding = 0x%" PRIx32
, s
->padding
);
1925 /* TODO: Blend shaders routing/disasm */
1926 union midgard_blend blend
= s
->blend
;
1927 mali_ptr shader
= pandecode_midgard_blend(&blend
, s
->unknown2_3
& MALI_HAS_BLEND_SHADER
);
1929 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
1931 pandecode_msg("mdg_blend = %" PRIx64
"\n", s
->blend
.shader
);
1935 pandecode_log("};\n");
1937 /* MRT blend fields are used whenever MFBD is used, with
1938 * per-RT descriptors */
1940 if (job_type
== MALI_JOB_TYPE_TILER
&& (is_bifrost
|| p
->shared_memory
& MALI_MFBD
)) {
1941 void* blend_base
= (void *) (s
+ 1);
1943 for (unsigned i
= 0; i
< fbd_info
.rt_count
; i
++) {
1944 mali_ptr shader
= 0;
1947 shader
= pandecode_bifrost_blend(blend_base
, job_no
, i
);
1949 shader
= pandecode_midgard_blend_mrt(blend_base
, job_no
, i
);
1952 pandecode_blend_shader_disassemble(shader
, job_no
, job_type
, false, gpu_id
);
1957 pandecode_msg("XXX: missing shader descriptor\n");
1960 DUMP_ADDR("Viewport", VIEWPORT
, p
->viewport
, 1);
1962 unsigned max_attr_index
= 0;
1964 if (p
->attribute_meta
)
1965 max_attr_index
= pandecode_attribute_meta(job_no
, attribute_count
, p
, false, suffix
);
1967 if (p
->attributes
) {
1968 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->attributes
);
1969 pandecode_attributes(attr_mem
, p
->attributes
, job_no
, suffix
, max_attr_index
, false, job_type
);
1972 /* Varyings are encoded like attributes but not actually sent; we just
1973 * pass a zero buffer with the right stride/size set, (or whatever)
1974 * since the GPU will write to it itself */
1976 if (p
->varying_meta
) {
1977 varying_count
= pandecode_attribute_meta(job_no
, varying_count
, p
, true, suffix
);
1981 attr_mem
= pandecode_find_mapped_gpu_mem_containing(p
->varyings
);
1983 /* Number of descriptors depends on whether there are
1984 * non-internal varyings */
1986 pandecode_attributes(attr_mem
, p
->varyings
, job_no
, suffix
, varying_count
, true, job_type
);
1989 if (p
->uniform_buffers
) {
1990 if (uniform_buffer_count
)
1991 pandecode_uniform_buffers(p
->uniform_buffers
, uniform_buffer_count
, job_no
);
1993 pandecode_msg("warn: UBOs specified but not referenced\n");
1994 } else if (uniform_buffer_count
)
1995 pandecode_msg("XXX: UBOs referenced but not specified\n");
1997 /* We don't want to actually dump uniforms, but we do need to validate
1998 * that the counts we were given are sane */
2002 pandecode_uniforms(p
->uniforms
, uniform_count
);
2004 pandecode_msg("warn: Uniforms specified but not referenced\n");
2005 } else if (uniform_count
)
2006 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2009 pandecode_textures(p
->textures
, texture_count
, job_no
, is_bifrost
);
2011 if (p
->sampler_descriptor
)
2012 pandecode_samplers(p
->sampler_descriptor
, sampler_count
, job_no
, is_bifrost
);
2016 pandecode_gl_enables(uint32_t gl_enables
, int job_type
)
2018 pandecode_log(".gl_enables = ");
2020 pandecode_log_decoded_flags(gl_enable_flag_info
, gl_enables
);
2022 pandecode_log_cont(",\n");
2026 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix
*p
, int job_no
, bool is_bifrost
)
2028 if (p
->shader
& 0xF)
2029 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p
->shader
& 0xF));
2031 pandecode_log(".postfix = {\n");
2034 pandecode_gl_enables(p
->gl_enables
, MALI_JOB_TYPE_TILER
);
2035 pandecode_prop("instance_shift = 0x%x", p
->instance_shift
);
2036 pandecode_prop("instance_odd = 0x%x", p
->instance_odd
);
2039 pandecode_msg("XXX: vertex only zero tripped");
2040 pandecode_prop("zero4 = 0x%" PRIx32
, p
->zero4
);
2043 pandecode_prop("offset_start = 0x%x", p
->offset_start
);
2046 pandecode_msg("XXX: vertex only zero tripped");
2047 pandecode_prop("zero5 = 0x%" PRIx64
, p
->zero5
);
2050 MEMORY_PROP(p
, position_varying
);
2051 MEMORY_PROP(p
, occlusion_counter
);
2054 pandecode_log("},\n");
2058 pandecode_tiler_heap_meta(mali_ptr gpu_va
, int job_no
)
2060 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2061 const struct bifrost_tiler_heap_meta
*PANDECODE_PTR_VAR(h
, mem
, gpu_va
);
2063 pandecode_log("struct bifrost_tiler_heap_meta tiler_heap_meta_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
2067 pandecode_msg("XXX: tiler heap zero tripped\n");
2068 pandecode_prop("zero = 0x%x", h
->zero
);
2071 pandecode_prop("heap_size = 0x%x", h
->heap_size
);
2072 MEMORY_PROP(h
, tiler_heap_start
);
2073 MEMORY_PROP(h
, tiler_heap_free
);
2075 /* this might point to the beginning of another buffer, when it's
2076 * really the end of the tiler heap buffer, so we have to be careful
2077 * here. but for zero length, we need the same pointer.
2080 if (h
->tiler_heap_end
== h
->tiler_heap_start
) {
2081 MEMORY_PROP(h
, tiler_heap_start
);
2083 char *a
= pointer_as_memory_reference(h
->tiler_heap_end
- 1);
2084 pandecode_prop("tiler_heap_end = %s + 1", a
);
2088 for (int i
= 0; i
< 10; i
++) {
2089 if (h
->zeros
[i
] != 0) {
2090 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2095 if (h
->unk1
!= 0x1) {
2096 pandecode_msg("XXX: tiler heap unk1 tripped\n");
2097 pandecode_prop("unk1 = 0x%x", h
->unk1
);
2100 if (h
->unk7e007e
!= 0x7e007e) {
2101 pandecode_msg("XXX: tiler heap unk7e007e tripped\n");
2102 pandecode_prop("unk7e007e = 0x%x", h
->unk7e007e
);
2106 pandecode_log("};\n");
2110 pandecode_tiler_meta(mali_ptr gpu_va
, int job_no
)
2112 struct pandecode_mapped_memory
*mem
= pandecode_find_mapped_gpu_mem_containing(gpu_va
);
2113 const struct bifrost_tiler_meta
*PANDECODE_PTR_VAR(t
, mem
, gpu_va
);
2115 pandecode_tiler_heap_meta(t
->tiler_heap_meta
, job_no
);
2117 pandecode_log("struct bifrost_tiler_meta tiler_meta_%"PRIx64
"_%d = {\n", gpu_va
, job_no
);
2120 pandecode_prop("tiler_heap_next_start = 0x%" PRIx32
, t
->tiler_heap_next_start
);
2121 pandecode_prop("used_hierarchy_mask = 0x%" PRIx32
, t
->used_hierarchy_mask
);
2123 if (t
->hierarchy_mask
!= 0xa &&
2124 t
->hierarchy_mask
!= 0x14 &&
2125 t
->hierarchy_mask
!= 0x28 &&
2126 t
->hierarchy_mask
!= 0x50 &&
2127 t
->hierarchy_mask
!= 0xa0)
2128 pandecode_prop("XXX: Unexpected hierarchy_mask (not 0xa, 0x14, 0x28, 0x50 or 0xa0)!");
2130 pandecode_prop("hierarchy_mask = 0x%" PRIx16
, t
->hierarchy_mask
);
2132 pandecode_prop("flags = 0x%" PRIx16
, t
->flags
);
2134 pandecode_prop("width = MALI_POSITIVE(%d)", t
->width
+ 1);
2135 pandecode_prop("height = MALI_POSITIVE(%d)", t
->height
+ 1);
2138 pandecode_msg("XXX: tiler meta zero tripped\n");
2139 pandecode_prop("zero0 = 0x%" PRIx64
, t
->zero0
);
2142 for (int i
= 0; i
< 12; i
++) {
2143 if (t
->zeros
[i
] != 0) {
2144 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64
"\n",
2150 pandecode_log("};\n");
2154 pandecode_primitive_size(union midgard_primitive_size u
, bool constant
)
2156 if (u
.pointer
== 0x0)
2159 pandecode_log(".primitive_size = {\n");
2163 pandecode_prop("constant = %f", u
.constant
);
2165 MEMORY_PROP((&u
), pointer
);
2169 pandecode_log("},\n");
2173 pandecode_tiler_only_bfr(const struct bifrost_tiler_only
*t
, int job_no
)
2175 pandecode_log_cont("{\n");
2178 /* TODO: gl_PointSize on Bifrost */
2179 pandecode_primitive_size(t
->primitive_size
, true);
2181 if (t
->zero1
|| t
->zero2
|| t
->zero3
|| t
->zero4
|| t
->zero5
2183 pandecode_msg("XXX: tiler only zero tripped\n");
2184 pandecode_prop("zero1 = 0x%" PRIx64
, t
->zero1
);
2185 pandecode_prop("zero2 = 0x%" PRIx64
, t
->zero2
);
2186 pandecode_prop("zero3 = 0x%" PRIx64
, t
->zero3
);
2187 pandecode_prop("zero4 = 0x%" PRIx64
, t
->zero4
);
2188 pandecode_prop("zero5 = 0x%" PRIx64
, t
->zero5
);
2189 pandecode_prop("zero6 = 0x%" PRIx64
, t
->zero6
);
2193 pandecode_log("},\n");
2197 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header
*h
,
2198 const struct pandecode_mapped_memory
*mem
,
2199 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2201 struct bifrost_payload_vertex
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2203 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
2205 pandecode_log("struct bifrost_payload_vertex payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
2208 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, false);
2209 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, true);
2212 pandecode_log("};\n");
2218 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header
*h
,
2219 const struct pandecode_mapped_memory
*mem
,
2220 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2222 struct bifrost_payload_tiler
*PANDECODE_PTR_VAR(t
, mem
, payload
);
2224 pandecode_vertex_tiler_postfix_pre(&t
->postfix
, job_no
, h
->job_type
, "", true, gpu_id
);
2225 pandecode_tiler_meta(t
->tiler
.tiler_meta
, job_no
);
2227 pandecode_log("struct bifrost_payload_tiler payload_%"PRIx64
"_%d = {\n", payload
, job_no
);
2230 pandecode_vertex_tiler_prefix(&t
->prefix
, job_no
, false);
2232 pandecode_log(".tiler = ");
2233 pandecode_tiler_only_bfr(&t
->tiler
, job_no
);
2235 pandecode_vertex_tiler_postfix(&t
->postfix
, job_no
, true);
2238 pandecode_log("};\n");
2244 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header
*h
,
2245 const struct pandecode_mapped_memory
*mem
,
2246 mali_ptr payload
, int job_no
, unsigned gpu_id
)
2248 struct midgard_payload_vertex_tiler
*PANDECODE_PTR_VAR(v
, mem
, payload
);
2249 bool is_graphics
= (h
->job_type
== MALI_JOB_TYPE_VERTEX
) || (h
->job_type
== MALI_JOB_TYPE_TILER
);
2251 pandecode_vertex_tiler_postfix_pre(&v
->postfix
, job_no
, h
->job_type
, "", false, gpu_id
);
2253 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no
);
2256 pandecode_vertex_tiler_prefix(&v
->prefix
, job_no
, is_graphics
);
2257 pandecode_vertex_tiler_postfix(&v
->postfix
, job_no
, false);
2259 bool has_primitive_pointer
= v
->prefix
.unknown_draw
& MALI_DRAW_VARYING_SIZE
;
2260 pandecode_primitive_size(v
->primitive_size
, !has_primitive_pointer
);
2263 pandecode_log("};\n");
2269 pandecode_fragment_job(const struct pandecode_mapped_memory
*mem
,
2270 mali_ptr payload
, int job_no
,
2271 bool is_bifrost
, unsigned gpu_id
)
2273 const struct mali_payload_fragment
*PANDECODE_PTR_VAR(s
, mem
, payload
);
2275 bool is_mfbd
= s
->framebuffer
& MALI_MFBD
;
2277 if (!is_mfbd
&& is_bifrost
)
2278 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
2280 struct pandecode_fbd info
;
2283 info
= pandecode_mfbd_bfr(s
->framebuffer
& FBD_MASK
, job_no
, true, false, is_bifrost
);
2285 info
= pandecode_sfbd(s
->framebuffer
& FBD_MASK
, job_no
, true, gpu_id
);
2287 /* Compute the tag for the tagged pointer. This contains the type of
2288 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
2289 * additional structures follow the MFBD header (an extra payload or
2290 * not, as well as a count of render targets) */
2292 unsigned expected_tag
= is_mfbd
? MALI_MFBD
: 0;
2296 expected_tag
|= MALI_MFBD_TAG_EXTRA
;
2298 expected_tag
|= (MALI_POSITIVE(info
.rt_count
) << 2);
2301 if ((s
->min_tile_coord
| s
->max_tile_coord
) & ~(MALI_X_COORD_MASK
| MALI_Y_COORD_MASK
)) {
2302 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2303 pandecode_prop("min_tile_coord = 0x%X\n", s
->min_tile_coord
);
2304 pandecode_prop("max_tile_coord = 0x%X\n", s
->max_tile_coord
);
2307 /* Extract tile coordinates */
2309 unsigned min_x
= MALI_TILE_COORD_X(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
2310 unsigned min_y
= MALI_TILE_COORD_Y(s
->min_tile_coord
) << MALI_TILE_SHIFT
;
2312 unsigned max_x
= (MALI_TILE_COORD_X(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
2313 unsigned max_y
= (MALI_TILE_COORD_Y(s
->max_tile_coord
) + 1) << MALI_TILE_SHIFT
;
2315 /* For the max, we also want the floored (rather than ceiled) version for checking */
2317 unsigned max_x_f
= (MALI_TILE_COORD_X(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
2318 unsigned max_y_f
= (MALI_TILE_COORD_Y(s
->max_tile_coord
)) << MALI_TILE_SHIFT
;
2320 /* Validate the coordinates are well-ordered */
2323 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
2324 else if (min_x
> max_x
)
2325 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
2328 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x
, max_x
);
2329 else if (min_y
> max_y
)
2330 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x
, max_x
);
2332 /* Validate the coordinates fit inside the framebuffer. We use floor,
2333 * rather than ceil, for the max coordinates, since the tile
2334 * coordinates for something like an 800x600 framebuffer will actually
2335 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2337 if ((min_x
> info
.width
) || (max_x_f
> info
.width
))
2338 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2340 if ((min_y
> info
.height
) || (max_y_f
> info
.height
))
2341 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2343 /* After validation, we print */
2345 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x
, min_y
, max_x
, max_y
);
2347 /* The FBD is a tagged pointer */
2349 unsigned tag
= (s
->framebuffer
& ~FBD_MASK
);
2351 if (tag
!= expected_tag
)
2352 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag
, tag
);
2357 /* Entrypoint to start tracing. jc_gpu_va is the GPU address for the first job
2358 * in the chain; later jobs are found by walking the chain. Bifrost is, well,
2359 * if it's bifrost or not. GPU ID is the more finegrained ID (at some point, we
2360 * might wish to combine this with the bifrost parameter) because some details
2361 * are model-specific even within a particular architecture. Minimal traces
2362 * *only* examine the job descriptors, skipping printing entirely if there is
2363 * no faults, and only descends into the payload if there are faults. This is
2364 * useful for looking for faults without the overhead of invasive traces. */
2367 pandecode_jc(mali_ptr jc_gpu_va
, bool bifrost
, unsigned gpu_id
, bool minimal
)
2369 pandecode_dump_file_open();
2371 struct mali_job_descriptor_header
*h
;
2372 unsigned job_descriptor_number
= 0;
2375 struct pandecode_mapped_memory
*mem
=
2376 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va
);
2380 h
= PANDECODE_PTR(mem
, jc_gpu_va
, struct mali_job_descriptor_header
);
2382 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2383 * high 32-bits of the 64-bit pointer are reused to store
2386 int offset
= h
->job_descriptor_size
== MALI_JOB_32
&&
2387 h
->job_type
!= MALI_JOB_TYPE_FRAGMENT
? 4 : 0;
2388 mali_ptr payload_ptr
= jc_gpu_va
+ sizeof(*h
) - offset
;
2390 payload
= pandecode_fetch_gpu_mem(mem
, payload_ptr
, 256);
2392 int job_no
= job_descriptor_number
++;
2394 /* If the job is good to go, skip it in minimal mode */
2395 if (minimal
&& (h
->exception_status
== 0x0 || h
->exception_status
== 0x1))
2398 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64
"_%d = {\n", jc_gpu_va
, job_no
);
2401 pandecode_prop("job_type = %s", mali_job_type_as_str(h
->job_type
));
2403 if (h
->job_descriptor_size
)
2404 pandecode_prop("job_descriptor_size = %d", h
->job_descriptor_size
);
2406 if (h
->exception_status
&& h
->exception_status
!= 0x1)
2407 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2408 h
->exception_status
,
2409 (h
->exception_status
>> 16) & 0xFFFF,
2410 mali_exception_access_as_str((h
->exception_status
>> 8) & 0x3),
2411 h
->exception_status
& 0xFF);
2413 if (h
->first_incomplete_task
)
2414 pandecode_prop("first_incomplete_task = %d", h
->first_incomplete_task
);
2416 if (h
->fault_pointer
)
2417 pandecode_prop("fault_pointer = 0x%" PRIx64
, h
->fault_pointer
);
2420 pandecode_prop("job_barrier = %d", h
->job_barrier
);
2422 pandecode_prop("job_index = %d", h
->job_index
);
2424 if (h
->unknown_flags
)
2425 pandecode_prop("unknown_flags = %d", h
->unknown_flags
);
2427 if (h
->job_dependency_index_1
)
2428 pandecode_prop("job_dependency_index_1 = %d", h
->job_dependency_index_1
);
2430 if (h
->job_dependency_index_2
)
2431 pandecode_prop("job_dependency_index_2 = %d", h
->job_dependency_index_2
);
2434 pandecode_log("};\n");
2436 switch (h
->job_type
) {
2437 case MALI_JOB_TYPE_WRITE_VALUE
: {
2438 struct mali_payload_write_value
*s
= payload
;
2439 pandecode_log("struct mali_payload_write_value payload_%"PRIx64
"_%d = {\n", payload_ptr
, job_no
);
2441 MEMORY_PROP(s
, address
);
2443 if (s
->value_descriptor
!= MALI_WRITE_VALUE_ZERO
) {
2444 pandecode_msg("XXX: unknown value descriptor\n");
2445 pandecode_prop("value_descriptor = 0x%" PRIX32
, s
->value_descriptor
);
2449 pandecode_msg("XXX: set value tripped\n");
2450 pandecode_prop("reserved = 0x%" PRIX32
, s
->reserved
);
2453 pandecode_prop("immediate = 0x%" PRIX64
, s
->immediate
);
2455 pandecode_log("};\n");
2460 case MALI_JOB_TYPE_TILER
:
2461 case MALI_JOB_TYPE_VERTEX
:
2462 case MALI_JOB_TYPE_COMPUTE
:
2464 if (h
->job_type
== MALI_JOB_TYPE_TILER
)
2465 pandecode_tiler_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2467 pandecode_vertex_job_bfr(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2469 pandecode_vertex_or_tiler_job_mdg(h
, mem
, payload_ptr
, job_no
, gpu_id
);
2473 case MALI_JOB_TYPE_FRAGMENT
:
2474 pandecode_fragment_job(mem
, payload_ptr
, job_no
, bifrost
, gpu_id
);
2480 } while ((jc_gpu_va
= h
->next_job
));
2482 pandecode_map_read_write();